eagle0504/sec-13f-holdings

SEC Form 13F Hedge Fund Holdings

Quarterly US-listed equity holdings for 9 institutional managers, reconstructed from their own Form 13F-HR filings with the SEC.

Built for the trackers at y-yin.io/research and published here because the filings are public domain and the parsing is fiddly enough to be worth sharing. Read straight from each filing's infotable.xml — the structured document EDGAR renders its own filing pages from — so no HTML is scraped.

• Coverage: 9 funds, 452 fund-quarters, 194,682 position rows
• Period range: 2013-06-30 to 2026-06-30
• Completeness: every position reported in each filing is included
• Generated: 2026-08-24

Funds

Fund	CIK	Quarters	Range
AQR Capital Management LLC	0001167557	53	2013-06-30 → 2026-06-30
ARK Investment Management LLC	0001697748	39	2016-12-31 → 2026-06-30
Berkshire Hathaway Inc	0001067983	53	2013-06-30 → 2026-06-30
Bridgewater Associates, LP	0001350694	53	2013-06-30 → 2026-06-30
Greenlight Capital Inc	0001079114	43	2013-06-30 → 2023-12-31
Lone Pine Capital LLC	0001061165	53	2013-06-30 → 2026-06-30
Paulson & Co. Inc.	0001035674	53	2013-06-30 → 2026-06-30
Pershing Square Capital Management, L.P.	0001336528	52	2013-06-30 → 2026-03-31
Third Point LLC	0001040273	53	2013-06-30 → 2026-06-30

Configs

holdings (default) — one row per fund / quarter / security

column	meaning
fund_slug, fund_name, cik	the filing manager
period	quarter end the filing reports (YYYY-MM-DD)
quarter	same, as 2026Q2
filed, form	filing date and form type of the operative filing
cusip, issuer, class	the security
put_call	Put, Call, or empty for the underlying
share_type	SH shares, or PRN principal amount
shares, value_usd	position size as reported, in whole dollars
weight	share of that quarter's reported book
shares_prev, d_shares, d_shares_pct	change vs the prior filing
d_value_pct	dollar change, for comparison — see the warning below
action	NEW, ADD, TRIM, HOLD, EXIT, OTHER
is_other	true for the pooled remainder of a truncated book

Exits are materialised as explicit rows (action = EXIT, shares = 0) in the quarter the position disappeared, so a position being closed is queryable rather than an absence you have to infer.

quarters — one row per fund / quarter

Book-level summary: positions, book_usd, top5_weight, top10_weight, hhi, effective_n (= 1/HHI), plus provenance (accession, value_rescaled, filings_merged, truncated_to).

from datasets import load_dataset

holdings = load_dataset("eagle0504/sec-13f-holdings", "holdings", split="train")
quarters = load_dataset("eagle0504/sec-13f-holdings", "quarters", split="train")

Measure change in share count, not dollars

Dollar value conflates the manager's trading with the market's price move, and the two routinely disagree in sign. Third Point's 2026-Q2 filing shows Amazon at −9.8% shares but +3.2% value — they were selling into an appreciating position. Use d_shares_pct to answer "did they buy or sell"; d_value_pct is provided for comparison, not for classification. The action column is derived from share count.

Four corrections applied to the raw filings

Each of these fails silently if you parse EDGAR naively.

1. <value> changed units mid-history. The SEC's 2022 amendments switched the field from thousands of dollars to whole dollars starting with the 2022-12-31 period. Unadjusted, a book appears to jump 1000× in one quarter. Every earlier filing is rescaled, with the era detected from the median implied share price (value / shares) rather than the date alone. quarters.value_rescaled flags the rescaled rows.
2. A 13F-HR/A is not always a replacement. Its cover page declares an amendmentType: RESTATEMENT supersedes the original filing, but NEW HOLDINGS adds rows withheld earlier, typically under confidential treatment. Treating the latter as a replacement deletes most of the book — Berkshire's 2025-03-31 amendment carries 4 rows against an original of 110, which turns a $259.8B quarter into a $1.1B one. Amendments are merged or replaced according to their declared type; quarters.filings_merged counts the filings behind each row.
3. The holdings filename varies by era — infotable.xml, form13fInfoTable.xml, and per-filer variants — so it is discovered from each filing's index.json rather than assumed.
4. Pre-2013 filings are fixed-width plain text, not XML, with per-filer column layouts. They are excluded; coverage starts at 2013-06-30.

What a 13F does not contain

Long positions in US-listed equities and certain convertibles and options, as of the last day of the quarter, filed up to 45 days later. Short positions, cash, credit, non-US listings, and anything opened and closed inside a quarter never appear. A 13F is a lagged, partial snapshot, not a portfolio.

Manager-specific caveats worth knowing:

• Bridgewater expresses most of its macro views in futures, swaps, currencies and sovereign bonds, none of which a 13F captures; the reported equity book is dominated by broad index ETFs. Treat it as a sleeve, not the portfolio.
• AQR is systematic and files tens of thousands of line items — 21,077 for 2026-06-30, aggregating to 3,900 distinct securities once rows split across managers and discretion types are summed — where no individual position is a thesis.
• ARK holds through ETFs whose positions are already disclosed daily, so the 13F adds little beyond a consistent long history.
• Greenlight has filed no holdings report since 2023-12-31; its rows end there.
• Pershing Square filed a 13F-NT (notice) for 2026-06-30, reporting its holdings on another manager's filing, so its latest holdings report is 2026-03-31.

How the dataset is built

data.sec.gov/submissions/CIK##########.json     every filing the manager ever made
        |  filter to 13F-HR and 13F-HR/A, dedupe by accession number
        v
www.sec.gov/Archives/edgar/data/<cik>/<acc>/index.json   that filing's file list
        |  find the holdings XML (the name varies by era)
        v
.../infotable.xml                               the holdings table itself
        |  parse, rescale units, merge amendments, aggregate by security
        v
one nested JSON panel per fund
        |  flatten, compute per-position deltas, materialise exits
        v
holdings.csv + quarters.csv  (this dataset)

infotable.xml is the structured document EDGAR renders its own filing pages from, so nothing here is scraped out of HTML. The scripts are published under scripts/ in this repo; see scripts/README.md to regenerate the whole thing from EDGAR yourself.

Cleaning rules, in the order applied

1. Select the filings. Keep 13F-HR and 13F-HR/A for the CIK. EDGAR pages long histories across several JSON files; these are concatenated and deduplicated by accession number, so no filing can be applied twice.
2. Drop the pre-XML era. Filings before the 2013-06-30 period are fixed-width plain text with per-filer column layouts. Parsing them generically is unreliable, so they are excluded rather than guessed at. Coverage therefore starts at 2013-06-30 (2016-12-31 for ARK, whose first filing is later).
3. Locate the holdings table. Read from the filing's own index.json, since the filename varies by era (infotable.xml, form13fInfoTable.xml, and per-filer variants). Namespace prefixes differ too, so elements are matched on local name.
4. Rescale the value column. The SEC's 2022 amendments switched <value> from thousands of dollars to whole dollars at the 2022-12-31 period. Rather than trust the date, each filing's era is inferred from the median implied share price (value / shares): below $1 across a whole equity book means thousands. Rescaled quarters are flagged in quarters.value_rescaled. Everything in this dataset is whole dollars.
5. Resolve amendments. Filings for the same period are applied oldest to newest. A 13F-HR/A reads its amendmentType from its cover page: RESTATEMENT replaces what came before, NEW HOLDINGS is merged on top. quarters.filings_merged records how many filings back a row.
6. Aggregate to one row per security. A single issuer often appears on several lines — split across included managers or investment-discretion types. These are summed by (cusip, put_call). This is why AQR's 2026-06-30 filing has 21,077 line items but 3,900 rows here.
7. Compute concentration on the complete book — positions, top5_weight, top10_weight, hhi — before any truncation, so the figures describe the filing rather than a subset.
8. Flatten and derive. One row per (fund, quarter, security), with d_shares, d_shares_pct and d_value_pct against the prior filing, and an action derived from share count (ADD/TRIM past a ±2% band, else HOLD).
9. Materialise exits. A position present last quarter and absent now becomes an explicit row with action = EXIT and shares = 0, so closures are queryable instead of being an absence you have to infer.

Steps 1, 4, 5 and 6 are each a place where a naive parse produces confident-looking wrong numbers rather than an error.

Known gaps and honest limitations

• Coverage starts in 2013, so pre-crisis and crisis-era positioning is absent even though these managers were filing since the late 1990s.
• No sector or ticker. A 13F identifies securities by CUSIP only. Mapping CUSIP to ticker or sector needs an external source and is not attempted here; issuer is the name as filed, which is inconsistently spelled across eras (FACEBOOK INC and META PLATFORMS INC are the same company).
• weight is share of the reported book, not of the manager's assets. For managers whose real exposure is mostly outside 13F scope — Bridgewater above all — it is not a portfolio weight in any meaningful sense.
• Issuer names are not normalised across corporate actions, and CUSIPs change on reincorporation, so a continuous economic position can appear as an exit and a new position. Longitudinal work should expect this.
• action uses a fixed ±2% band on share count. It is a convenience, not a judgement; d_shares_pct is there if you want your own threshold.
• Amendments are resolved as filed. If a manager mis-declares an amendment type, this dataset inherits that error.

Provenance and licence

Source: SEC EDGAR, data.sec.gov and www.sec.gov/Archives. SEC filings are works of the US government and are in the public domain; this derived dataset is released under CC0. It is generated by industry-reports/13f/ in yiqiao-yin/y-yin-homepage — fetch_13f.py pulls and corrects the filings, push_to_hf.py flattens and publishes them.

Informational only. Not investment advice.